-
Notifications
You must be signed in to change notification settings - Fork 320
Pull requests: unslothai/unsloth-zoo
Author
Label
Projects
Milestones
Reviews
Assignee
Sort
Pull requests list
Read only the norm class when picking the layernorm precision
#1194
opened Sep 12, 2026 by
vineethsaivs
Loading…
Report the chunked prefill budget vLLM was actually given
#1193
opened Sep 12, 2026 by
vineethsaivs
Loading…
Make the undefined-name lint gate, and fix the two real bugs it was reporting
#1192
opened Sep 12, 2026 by
danielhanchen
Member
Loading…
perf(mlx): fuse recurrent decode convolution and SiLU
new-feature#
non-studio#
refactor#
#1191
opened Sep 10, 2026 by
Lyxot
Contributor
Loading…
perf(mlx): fuse quantized MoE gate and up projections
new-feature#
non-studio#
regresses#
Risks breaking or slowing training, inference, install, or changes the inference system prompt
#1189
opened Sep 10, 2026 by
Lyxot
Contributor
Loading…
fix(mlx): quantize a merged_4bit export whose base was never quantized
important#
non-studio#
#1186
opened Sep 9, 2026 by
Lyxot
Contributor
Loading…
fix(mlx): honor processor tokenizer trust and clarify refusals
important#
non-studio#
regresses#
Risks breaking or slowing training, inference, install, or changes the inference system prompt
#1185
opened Sep 9, 2026 by
Lyxot
Contributor
Loading…
fix(mlx): preserve VLM processor and training-path contracts
important#
non-studio#
refactor#
#1184
opened Sep 9, 2026 by
Lyxot
Contributor
Loading…
Build(deps): bump the python group with 7 updates
dependencies
Pull requests that update a dependency file
non-studio#
python
Pull requests that update python code
regresses#
Risks breaking or slowing training, inference, install, or changes the inference system prompt
#1182
opened Sep 9, 2026 by
dependabot
Bot
Loading…
fix(mlx): complete text-model routing and cache-free training
conflicts#
edge case#
non-studio#
refactor#
regresses#
Risks breaking or slowing training, inference, install, or changes the inference system prompt
#1178
opened Sep 8, 2026 by
Lyxot
Contributor
Loading…
MLX: bring DPO's reference policy to parity with TRL's DPOTrainer
new-feature#
regresses#
Risks breaking or slowing training, inference, install, or changes the inference system prompt
#1171
opened Sep 7, 2026 by
Lyxot
Contributor
Loading…
Fix the flex-attention backward shared-memory gate missing every num_stages=1 GPU
edge case#
#1165
opened Sep 6, 2026 by
danielhanchen
Member
Loading…
Do not report unsloth's own bitsandbytes stub as a test module leak
non-studio#
#1133
opened Aug 29, 2026 by
vineethsaivs
Loading…
Never hand a packed 4-bit weight to the Linear4bit F.linear fallback
edge case#
non-studio#
#1132
opened Aug 29, 2026 by
lonexreb
Loading…
MLX: int8 W8A8 prefill for quantized models on Apple M5
edge case#
mlx
new-feature#
refactor#
#1098
opened Aug 25, 2026 by
danielhanchen
Member
Loading…
Drop the vision and audio tensors from a text_only merged_16bit export
edge case#
new-feature#
non-studio#
#1097
opened Aug 25, 2026 by
Zelys-DFKH
Loading…
Fix text-only loading for VLM-only MLX models
edge case#
#1096
opened Aug 24, 2026 by
Supersynergy
Contributor
Loading…
fix: forward enable_thinking=False in vision collator for Qwen3.8
edge case#
#1072
opened Aug 18, 2026 by
Souravrajvi0
Loading…
2 tasks done
MLX VLM: name the config key when a mirrored config.json is incomplete
conflicts#
edge case#
refactor#
#1013
opened Aug 8, 2026 by
BardiaKoopah
Contributor
Loading…
MLX: name the config key when a mirrored config.json is incomplete
edge case#
#1004
opened Aug 6, 2026 by
BardiaKoopah
Contributor
Loading…
Feat(mlx): Support group_by_length for text training
conflicts#
edge case#
new-feature#
refactor#
#995
opened Aug 4, 2026 by
BardiaKoopah
Contributor
Loading…
Fix Qwen3.5 fp16 training dtype mismatches under UNSLOTH_FORCE_FLOAT32
edge case#
#978
opened Aug 4, 2026 by
chakshu-dhannawat
Loading…
Feat(mlx): Honor adam_epsilon in MLXTrainer
conflicts#
edge case#
new-feature#
#977
opened Aug 4, 2026 by
BardiaKoopah
Contributor
Loading…
Feat(mlx): GGUF multi-quant
conflicts#
edge case#
new-feature#
#976
opened Aug 4, 2026 by
BardiaKoopah
Contributor
Loading…
fix(mlx): make save_method='merged_4bit' actually quantize an unquantized model
important#
#971
opened Jul 31, 2026 by
BardiaKoopah
Contributor
Loading…
Previous Next
ProTip!
Find all pull requests that aren't related to any open issues with -linked:issue.